Grand Challenges in Ocean and Climate Modelling - STFC's ...

Grand Challenges in Ocean
and Climate Modelling
Jeff Blundell, Beverly De Cuevas, Adrian New
Bob Marsh, Katya Popova, Peter Challenor

Challenge 1: Mixing and Internal Waves
The realistic representation of mixing in ocean models is perhaps
the biggest challenge facing physical oceanographers today
1 cm 2 s -1
0.1 cm 2 s -1
This plot of zonal-mean vertical diffusivity (mixing rate) from the OCCAM 1/12° model
shows large diffusivities in the deep ocean, but the model does not resolve key processes,
in particular internal waves.

Internal Wave Processes
Internal waves are generated through the interaction of surface tides (from the
sun and moon) with topography, particularly steep shelf-slope
topography and “rough” mid-ocean ridge topography in the deep ocean.
Internal waves are an important mixing process in the ocean but occur with
length scales of typically a few km, so cannot be explicitly resolved in
today’s ocean models (i.e. the OCCAM 1/12° model has a resolution of
order 10km horizontally).

Shelf Slope Generated Internal Waves (Bay of Biscay)
Displacements
Horizontal
velocity
Internal tides are order 30-50 km
wavelength and are generated
near the shelf break. They break
down into shorter 1-2 km internal
waves shown here, through
nonlinear processes.
Vertical
velocity

Mixing near the Shelf Break (Bay of Biscay)
Surface chlorophyll on 4.9.99
The internal tides and waves
produce mixing of the upper
ocean near the shelf break.
This injects nutrients into the
well-lit layers and allows
plant cells (phytoplankton/
chlorophyll) to grow or
“bloom” near the shelf break
(200 m contour). So internal
wave mixing is important for
biological growth in the
ocean.

Internal Wave Mixing in the Deep Ocean
Internal waves are thought to
provide an important mixing mechanism
for stirring the deep ocean and upwelling
the deep waters to nearer the surface.
This is an important route whereby waters
which have left the surface can “overturn”
and once more return to the surface. The
mixing is therefore critical to maintaining the
large-scale circulation in the ocean.
The figure (from theory) shows that internal
tides and waves provide 0.9 TW of energy
to the deep ocean, nearly half of the total
required (the remainder being provided by
the wind).
From Munk & Wunsch (1998)

Example from the Indian Ocean: Mascarene Plateau
The topography west of the
Plateau is relatively smooth
and flat and does not produce
internal waves.
East of the Plateau the
seafloor is highly canyonated
due to the central Indian Ridge.
The canyons (features running
NE to SW) are typically
10-20 km across.

Mean Temperature Profiles
The mean temperature profile East of the Plateau looks like a “mixed” version of the
profile from West of the Plateau - as if internal waves were mixing the deeper colder
waters with lighter waters from above - upwelling them several thousand metres.

Inferred Vertical Diffusivities (Mixing Rates)
Method of Polzin (1995) using ‘strain’ (small scale variations in the vertical density gradient)
inferred from CTD measurements shows that mixing is actually much
higher East of the Plateau than West of it.
Central
West
East
K ρ ~ 10 -5 -10 -2
m 2 s -1

Diffusivities over “Rough” Topography East of Mascarene Plateau
Central
West
East
In particular, diffusivities
are much higher
in topographic troughs or
canyons than on ridges/
crests.
(The plots show mixing
estimates from 5 or 6
“dips” with the CTD at each
of two stations).
Note the roughness of the
topography - typical scales
of 10-20 km.

Internal Waves at Station 17 (Trough)
Internal waves of tidal period (and likely wavelengths of 10-20 km) are
found above the rough topography (right-hand figure) and appear to be
breaking in the canyons/ troughs (below 3500 m). The breaking of the waves is
causing the mixing of the deep ocean, and the upwelling of the deep waters.

NEMO 1/4° global ocean model surface temperature
Thus a global ocean model of order 1 km horizontal resolution would explicitly
resolve a good part of the internal wave spectrum. The resulting mixing could then
be easily included through (for example) a Richardson number dependent scheme.
Improved parameterisations could also be developed for lower resolution models.
This would be likely to cause a major change in the deep circulation patterns in
the global ocean, and in the upwelling of the deep waters, and in the way in which
the ocean circulation is maintained. Such a model would therefore lead to major
improvements in ocean and climate modelling systems and predictions, and to
our understanding of how the ocean works.

Challenge 2: Global Carbon Cycle
MesoZoo
MicroZoo
Diatoms
NanoPhyto
Si
Fe
N
Slow
detritus
Fast particle
flux
Medusa: Model of Ecosystem Dynamics,
nutrient Utilisation and SequestrAtion
Chlorophyll from Medusa ecosystem model
in global NEMO 1/4° model
Models of upper ocean biology (eg the “Medusa” model at NOCS) can be easily built into
standard ocean models, eg the 1/4° NEMO model above. Upper ocean biology quickly comes
into a near-equilibrium, after less than a decade at most. It drives a carbon sinking flux or
“pump” to the deeper ocean (a simple modification to Medusa would be needed for this),
removing it from the atmosphere. However, it would take order 1000 years for the carbon in the
deep ocean to come into equilibrium (a typical “overturning” timescale for the ocean).
Equilibration of the ocean carbon system is needed for a pre-industrial atmospheric CO2 level
before studies can be made of how this “carbon pump” will change in the future and its role in
slowing climate change. To date, this has only been achieved in models with 2° resolution.
However, a 1/4° or 1/12° resolution is needed to adequately model the interactions of eddies
(which affect mixed layer depth and upwelling of nutrients) with the biology.

Challenge 3: Climate Uncertainty
AMOC simulated with GENIE-1
(Marsh et al., 2007)
How the climate will change in the future is far from certain. For a particular
model, different outcomes will be obtained for different parameter settings in
the model (“parameter uncertainty”). The figure shows how the Atlantic
Meridional Overturning Circulation (AMOC) might change in the future (for a
given warming scenario) in the GENIE model, depending on how various
parameters in the model are chosen. To properly assess the model’s
response, about 10,000 – 100,000 runs would be needed to span the
parameter space using naïve Monte Carlo methods.

Challenge 3: Climate Uncertainty
Future AMOC simulated with
GENIE-1
(Challenor et al., 2008)
If we build an “emulator” (a statistical approximation to the full climate model) we can reduce this number of
runs to 0(100 - 1000) we can now apply statistical techniques, so that the probability of the collapse of the
AMOC (for instance) can be derived. (Such methods are being developed in the “Managing Uncertainty in
Complex Models” project, to minimise the number of runs we do).
For GENIE, the probability of AMOC collapse (with drastic implications for cooling of European climate
in the future) was estimated at about 30% (Challenor et al., 2006). However, GENIE has simplified physics
(e.g. planetary geostrophic ocean) and a coarse resolution (3-10°, 8 ocean depth levels). The exercise should
now be repeated for higher resolution full-physics GCMs which include more of the relevant processes. Such
models have order 100 uncertain parameters so our training ensemble will need to be bigger.

Summary and Requirements
Challenge 1: Internal Wave Mixing
Needs global or basin-scale ocean model (physics only) of order
1 km resolution, run for periods of order 30-40 years.
Challenge 2: Global Carbon Cycle
Needs global physics and biology ocean model, at highest achievable
resolution (1/4 or 1/12°) run for periods of order 1,000 years or more.
Challenge 3: Climate Uncertainty
Needs coupled climate models at best possible resolution (e.g. 1 or 0.5°)
run for periods of order 100-1000 years, but up to 1,000 times (or more).
These challenges clearly have very different requirements in terms of
number of points per cpu node, amounts of data exchange between
nodes, and amounts of IO etc. Also, critically, each challenge needs
significant extra manpower, which is currently lacking.
Some may be more easily achievable than others - Discussion needed.

Practical Issues:
1: Storage
a) Performance – may need fast I/O for adequate model performance,
so likely to want some specialised disk during run
b) Volume – Env. science has a particular requirement for large data volumes
as an example, current ¼ degree NEMO requires (for 50 year run)
7TBytes of raw output, and 18Tbytes of processed output for analysis
Extrapolating from this:
Challenge 1: 1,000 times larger -> 10's of Petabytes
Challenge 2: 20 x longer, 3 x larger for extra fields -> 60 times larger
Challenge 3: 1,000 runs, maybe some savings -> several Petabytes
So, Petascale means Petabytes as well as Petaflops
c) Archive: want readily-accessible archive for the above data volumes,
for a period of ~ 10 years (5 years of model running + 5 years ongoing analysis)

2: Computational environment
Practical Issues:
a) Associated computing – will need to perform large amounts of post-processing
(e.g. stitching together output from adjacent processors,
plus computation of diagnostic quantities)
Code less parallel and demanding, so not favoured by
scheduling at major centres. But either need to do it there
(for good access to data) or move vast volumes around
JANET and have own significant HPC provision.
b) Need own excellent infrastructure (much more widespread HPC)
for analysis and visualisation
c) Data serving – need to make post-processed data available to user community
data portals served by major centres, or by home institute?
d) However organised, will involve massive transfer, and require excellent
networking to all relevant sites

Practical Issues:
3: People
a) Staff to run project – funding not guaranteed, lousy career structure,
HPC skills appreciated in e.g. financial modelling
b) Expert assistance at major HPC sites:
knowledge of systems quirks, tuning tools etc.
Need specialised help with e.g. code scaling
c) User community – need HPC literate community of users to do the related
science. They are starting with ~no programming
knowledge, but will need to use powerful // systems
to make any progress with the data
d) Trend in recent HPC projects to significant resource allocation for people

Practical Issues:
Summary:
No lack of science ideas
(haven’t even discussed coupled modelling)
Pleased to see ambitions described as Petascale, not just Petaflop
Many practical problems (even today) are not with the compute engine,
but almost every other aspect of such projects.